ssonong / are-gazeestimation

Code for ECCV 2018, Appearance-Based Gaze Estimation via Evaluation-Guided Asymmetric Regression

Asymmetric Gaze Regression

Introduction

This is the README file for the official code associated with the ECCV2018 paper, "Appearance-Based Gaze Estimation via Evaluation-Guided Asymmetric Regression".

Our academic paper which describe ARE-Net in detail and provides full result can be found here: [PAPER].

Usage

We also ask that you cite the associated paper if you make use of this code; following is the BibTeX entry:

@inproceedings{eccv2018_are,
Author = {Yihua Cheng and Feng Lu and Xucong Zhang},
Title = {Appearance-Based Gaze Estimation via Evaluation-Guided Asymmetric Regression},
Year = {2018},
Booktitle = {European Conference on Computer Vision (ECCV)}
}

Environment

To using this code, you should make sure following libraries are installed first.

Python>=3
Tensorflow-GPU>=1.10
PyYAML==5.1
numpy, os, math etc., which can be found in the head of code.

Run the code

Config

You need to modify the config.yaml first especially data/label and data/root params.

data/label represents the path of label file.

data/root represents the path of image file.

A example of label file is data folder. Each line in label file is conducted as:

p00/left/1.bmp p00/right/1.bmp p00/day08/0069.bmp -0.244513310176,0.0520949295694,-0.968245505778 ... ...

Where our code reads image data form os.path.join(data/root, "p00/left/1.bmp") and reads gts of gaze direction from the rest in label file.

Options

We provide two optional args, which are -m and -n.

-m represet the running mode. We use 1 for train mode, 2 for predict mode and 3 for evaluate mode.

For example, you can run the code like:

python main.py -m 13

to train and evaluate model together.

-n represet the number of test file in 'leave-one-person-out' strategy.
For example, data/label provide 15 label file. Use

python main.py -m 13 -n 0

, you train and evaluate the model with using the first person (p00.label) as test file.
Note that, we add a loop in main.py to perform leave-one-person-out automatically. You can delete it for your individual usage.

Result